On the architecture axis.
Our architecture assessment was not an inference from resemblance. It rested on counts that anyone can reproduce from the published files, and we think it is worth stating them precisely.
Against the 94,172 tensors listed in A.X-K2's model.safetensors.index.json, tensor names are identical for 93,382 (99.2%) with Kimi-K2-Base, 90,721 (96.3%) with DeepSeek-V3.2-Exp, and 90,294 (95.9%) with DeepSeek-V3. At configuration level, A.X-K2 carries all twenty DeepSeek-family-specific keys β q_lora_rank, kv_lora_rank, qk_nope_head_dim, qk_rope_head_dim, v_head_dim, n_routed_experts, n_shared_experts, moe_intermediate_size, first_k_dense_replace, moe_layer_freq, n_group, topk_group, topk_method, scoring_func, routed_scaling_factor, norm_topk_prob, num_experts_per_tok, index_head_dim, index_n_heads, index_topk β and shares fourteen non-obvious constants simultaneously with DeepSeek-V3, V3.2-Exp, Kimi-K2-Base and your own A.X-K1: hidden 7168, intermediate 18432, 61 layers, q_lora 1536, kv_lora 512, qk_nope 128, qk_rope 64, v_head 128, moe_intermediate 2048, 8 experts per token, 1 shared expert, sigmoid scoring, noaux_tc, rms_eps 1e-6. The sparse-attention indexer constants (index_head_dim 128, index_n_heads 64, index_topk 2048) match the DeepSeek V3.2 family exactly, as do the corresponding tensor shapes.
Where the label was wrong. "Exact-match" is not supportable and we are withdrawing it. Your figures are correct: 64 attention and KV heads against 128, one initial dense layer against three, vocabulary 163,840 against 129,280. At tensor level this shows up as o_proj [7168, 8192] against [7168, 16384] and kv_b_proj [16384, 512] against [32768, 512], and A.X-K2 additionally carries rank-16 gated-norm weights on all 61 layers with no counterpart in DeepSeek. A label of partial is the accurate one.
Where we do not think we were wrong. The measurements above describe substantial adoption of published design, and we read your filing as agreeing on this point rather than disputing it: the correction you request on this axis is foreign β partial, not foreign β self, and your model card credits DeepSeek-V3.2 for the sparse attention, Qwen for gated attention and gated norm, and Kimi K2 for the head-count finding. We will present the axis as partial adoption with per-component provenance rather than as a single-baseline match.
What this is not. Adopting a published architecture is standard practice and carries no implication of impropriety. For calibration: Kimi-K2-Base ships DeepSeek's own modeling_deepseek.py (75,769 bytes against DeepSeek's 75,741) and declares DeepseekV3ForCausalLM, and comparable routing choices β sigmoid scoring with noaux_tc balancing, routed_scaling_factor 2.5, 256 routed experts β appear in several models from unrelated organizations released over the same period. Design reuse at this scale is the field's baseline, not a deviation from it. Questions of training independence belong to the weights axis, which we treat separately and address below.
One correction to our own reading. We had attributed A.X-K2's 64 heads, 163,840 vocabulary and single dense layer to Kimi-K2. That is wrong: A.X-K1 (2025-12-29) already carries all of them, along with 61 layers, hidden 7168, q_lora_rank 1536 and routed_scaling_factor 2.5. They are your own prior generation's settings. We also note that the low tensor-name overlap between A.X-K1 and A.X-K2 (0.7%) is an artifact of storage layout β A.X-K1 stores experts stacked per layer β and not a lineage signal; we mention it so the figure is not misread in either direction.
